University of Cambridge
2022-2025
Brief Description
This PhD combines systems engineering with human-centred design to determine how digital twin representations can be made fit for purpose. The work spans the full stack: from low-level sensor design and real-time platform architecture to visualisation creation and end-user evaluation.
Using 39 sensors, 21 visualisations, and two novel frameworks (DiTTo and PUX), five empirical studies with real users produced practical, evidence-based guidance on which DT representations work best for different activities.
This PhD thesis investigates digital twins from both a systems engineering and human-centred design perspective: how can we ensure that digital twin representations are fit for purpose?
Using Lecture Theatre 1 (LT1) in the Cambridge Computer Laboratory as a testbed, I deployed 39 sensors (36 Elsys environmental sensors and 3 Cerberus privacy-preserving cameras) connected to the Adaptive City Platform (ACP) for real-time data processing. From this infrastructure, I created 21 digital twin visualisations spanning different levels of fidelity, timeliness, and data aggregation.
To evaluate these visualisations, I developed two frameworks: DiTTo (Digital Twin Taxonomy), which classifies DT representations along three design axes, and a DT-adapted version of Blackwell's Patterns of User Experience (PUX), which maps user activities to suitable visualisation types. Five empirical studies with three actor groups (Facility Managers, DT Practitioners, and Building Occupants) validated these frameworks and revealed that activity type predicts visualisation preference better than actor identity. The thesis addresses three research questions:
- RQ1 - Real-time tracking and event detection
- RQ2 - Visual fidelity
- RQ3 - Predicting design criteria from activities
The dissertation stack below illustrates how the research was conducted, from low-level sensor design and real-time platform development, through visualisation design, all the way to end-user testing.
Data Visualisations
I created 21 unique DT visualisations based on dense sensor deployments in LT1, designed to evaluate which representations best support the activities of different user groups. For an in-depth breakdown of visualisation classification and design rationale, see the DiTTo page.
System Architecture
The platform is built around a low-latency, asynchronous architecture. Over 400 IoT sensors (~2M messages/day) feed into an async pub-sub data pipeline with reliable ingestion and routing via the Adaptive City Platform (ACP). Event-detection models run on edge devices with end-to-end latency under 100 ms and stable throughput, using ontology-based data models for consistent sensor representation.
High-performance dashboards are powered by custom APIs and WebSocket data streams, driving interactive D3.js and Unity visualisation layers. The full system supports real-time in-building monitoring and derived intelligence across all 21 DT representations.
Research Questions
RQ1:
How can we design digital twins for real-time tracking and event detection?
The Adaptive City Platform (ACP) processes sensor inputs with sub-100 ms latency via an event-driven architecture, enabling live synchronisation between physical events and their digital representations. The system distinguishes between simple events (raw sensor readings) and derived events (computed from multiple inputs, such as occupancy counts or anomaly flags). For privacy-preserving crowd sensing, the Cerberus system performs ensemble face detection on-device using Raspberry Pi 4, transmitting only aggregate counts. Real-time visualisations, such as animated ripple sequences representing motion through corridors, make detected events immediately interpretable. The "synopsis" concept aggregates event histories into compact visual summaries, allowing users to review building activity without replaying raw data streams.
RQ2:
What is the appropriate level of visual fidelity for digital-twin representations?
The DiTTo (Digital Twin Taxonomy) framework classifies DT representations along three axes: timeliness, fidelity, and aggregation. User studies revealed that 2D abstract visualisations were preferred for high-timeliness expert tasks such as real-time monitoring and fault detection, where speed and clarity outweigh immersion. In contrast, 3D and higher-fidelity representations were valued in social and communicative contexts: stakeholder presentations, collaborative planning, and spatial understanding. The appropriate fidelity level is therefore not a fixed property of the system but a function of the activity it supports.
RQ3:
Can we predict design criteria for digital twins based on the activities they will support?
Using Blackwell's Patterns of User Experience (PUX) framework, I mapped user activities to DT visualisation preferences through PUX Activity Scores. Across five empirical studies with Facility Managers, DT Practitioners, and Building Occupants, five out of eight tested activities showed significant positive correlations between predicted and actual visualisation preferences. Crucially, activity type was a stronger predictor of visualisation suitability than actor identity - meaning what users need to do matters more than who they are when selecting a digital twin representation.
Thesis submitted November 2025, currently awaiting viva. The full text will be made public after examination.